CCE Theses and Dissertations

Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)


College of Computing and Engineering


Wei Li

Committee Member

Ling Wang

Committee Member

Ajoy Kumar


Android, computer security, feature selection, machine learning, malware detection


Malicious applications pose an enormous security threat to mobile computing devices. Currently 85% of all smartphones run Android, Google’s open-source operating system, making that platform the primary threat vector for malware attacks. Android is a platform that hosts roughly 99% of known malware to date, and is the focus of most research efforts in mobile malware detection due to its open source nature. One of the main tools used in this effort is supervised machine learning. While a decade of work has made a lot of progress in detection accuracy, there is an obstacle that each stream of research is forced to overcome, feature selection, i.e., determining which attributes of Android are most effective as inputs into machine learning models.

This dissertation aims to address that problem by providing the community with an exhaustive analysis of the three primary types of Android features used by researchers: Permissions, Intents and API Calls. The intent of the report is not to describe a best performing feature set or a best performing machine learning model, nor to explain why certain Permissions, Intents or API Calls get selected above others, but rather to provide a holistic methodology to help guide feature selection for Android malware detection.

The experiments used eleven different feature selection techniques covering filter methods, wrapper methods and embedded methods. Each feature selection technique was applied to seven different datasets based on the seven combinations available of Permissions, Intents and API Calls. Each of those seven datasets are from a base set of 119k Android apps. All of the result sets were then validated against three different machine learning models, Random Forest, SVM and a Neural Net, to test applicability across algorithm type.

The experiments show that using a combination of Permissions, Intents and API Calls produced higher accuracy than using any of those alone or in any other combination and that feature selection should be performed on the combined dataset, not by feature type and then combined. The data also shows that, in general, a feature set size of 200 or more attributes is required for optimal results. Finally, the feature selection methods Relief, Correlation-based Feature Selection (CFS) and Recursive Feature Elimination (RFE) using a Neural Net are not satisfactory approaches for Android malware detection work.

Based on the proposed methodology and experiments, this research provided insights into feature selection – a significant but often overlooked issue in Android malware detection. We believe the results reported herein is an important step for effective feature evaluation and selection in assisting malware detection especially for datasets with a large number of features. The methodology also has the potential to be applied to similar malware detection tasks or even in broader domains such as pattern recognition.